Making life harder for recruitment scammers. Does your family have a secret codeword?
Claude gets watermarks. US authorises cyber privateers. And the Musk effect.
Job security
Recruitment has long been a favoured way to attack organisations and scam individuals, with LinkedIn a key element in the chain. Among this week’s news: the FBI is investigating how an unnamed US federal agency ended up hiring a North Korean operative as a remote IT worker, North Korean hackers posing as recruiters targeted defence, aerospace and aviation workers with convincing job offers, and fake headhunters are pursuing employees at companies caught up in mergers and takeovers.
We call out LinkedIn because it provides attackers with a curated set of information ideally suited to their nefarious ends. By default, career history, education, and interests are open to anyone to view - and AI can harvest that information (even if mass “scraping” is against LinkedIn’s terms and conditions). In the case of one organisation in the process of a merger, we’ve seen employees receiving improbably attractive job offers in emails obviously based on their LinkedIn employment history. Our advice is to minimise the detail exposed publicly; not for nothing is LinkedIn known as the “phisherman’s friend.”
Threats
Fakers: A Hong Kong man lost US$1.27 million after scammers used AI to impersonate his father on WhatsApp. Police said the victim was repeatedly targeted until his savings were exhausted. While this type of fraud is particularly common in Hong Kong, where there have been 150 incidents in just two weeks resulting in losses of US$3.3 million, it’s a growing threat everywhere. It really is worth having a codeword that family and friends can use in the event of a genuine emergency.
Whack-a-mole: A malicious Chrome extension that was yanked from the Web Store in January is back under a new guise. And a researcher has found more than 700 free VPN extensions intercepting browser traffic. Honestly, avoid adding things to your browser unless you’re absolutely sure they’re safe.
Free: Criminals are creating convincing copies of popular websites, including CNN and Avast, with the aim of persuading Windows users to download “free” apps. In fact, the downloads install remote-management software linked to the attackers’ accounts, potentially giving them control of the computer. The software itself is legitimate, meaning antivirus may not catch it. If you need an app, download it from an official store - not via a link. Malwarebytes
Clones: Fraudsters are targeting Android users with convincing phone calls to steal money from contactless payment cards. Posing as bank staff, they persuade victims to install an app, tap their card against their phone and enter their PIN. Malicious software then relays the card interaction to the criminals, allowing them to make apparently legitimate payments or ATM withdrawals. Researchers have seen successful attacks completed during calls lasting just 13 minutes. We should never trust any calls claiming to be from a financial institution. If in doubt, call them on a number you know to be genuine. Group-IB
AI
Watermark: Anthropic has started adding a watermark to content generated by its Claude models, beginning with the latest ones. Although the move is in response to the EU AI Act, the change applies worldwide and covers anything processed by Claude, not just what it has authored. The invisible watermark applied to text exploits the way models work. When they write, they pick a word from a set of options. The watermarking process splits those options into two groups using a secret key and pushes the model towards one of them. Clearly, a single word proves nothing but over several hundred words the pattern emerges and can be detected (though Anthropic has yet to release a detector). Also worth noting; the web is now flooded with “watermark removers.” They can’t do anything about text-based watermarks for the moment, so are a waste of time and money.
Skills: AI skills are instructions that enable AI agents to perform specific tasks which obviously makes them powerful - and an attractive target for attackers. Zenity Labs has found a campaign that clones existing skills, gives them an almost identical name, and after waiting a while inserts malicious code designed to steal information. Zenity sells a product to secure AI agents, so it’s hardly a disinterested party, but its research does highlight an issue users need to be aware of.
Basics: Some valuable context from Resilience, which is a leading cyber insurer. “AI is already reshaping cyber risk, just not in the way most people expect. Rather than creating a new category of AI-native losses on its own, it’s making old attack methods, especially phishing and social engineering, dramatically more effective,” it says. According to its half-year report, zero losses resulted from an “AI-specific attack vector.” By contrast, losses linked to phishing, social engineering, and transfer fraud rose from 17.7% of the total in the first half of 2024 to 85.3% in the same period this year.
Zucking: A quick word about Mark Zuckerberg’s mammoth 6,500 word essay in which he argues (somewhat unconvincingly given Meta’s history) that superintelligence should be widely distributed rather than controlled by a few institutions. Meta’s track record in developing AI models has been patchy, to say the least. As Bloomberg Opinion pithily put it, “the only reason Meta isn’t talking up an all-powerful closed AI model for businesses, governments [and] institutions is that it has so far failed to build a competitive one.”
Mountebank*: A company offering “100% human-written, Never AI” medical research is entirely AI driven, according to 404 Media. Research Gold’s team of human methodologists are either AI-generated or using the identity of real people without their permission, it found.
*A person who tricks others by pretending to have skills, knowledge, or miracle cures they do not actually possess.
US
Privateers: In a major change to US government policy, vetted private companies will be allowed to launch offensive cyber operations against international criminal gangs and hackers. A presidential memorandum said the move would allow the federal government to use “innovative capabilities of the private sector” to combat cybercrime and threats targeting Americans. Companies wishing to take part will operate under federal oversight and will have to stump up a $1 million bond which they will forfeit if things go wrong.
Feed: Not-for-profit news groups are suing over the scheme to sell real-time access to President Trump’s Truth Social posts. “This scheme is profoundly corrupt. The President stands to gain financially by giving ‘market-moving’ government information to those who are willing and able to pay his personal company,” the lawsuit says.
Cameras: Surveillance technology provider Flock is releasing updates to address reports of police turning its tools into a personal stalking service. The changes mean a previously optional feature to flag up “abnormal search behaviour” will become mandatory. Civil rights groups are unconvinced. The ACLU said Flock’s changes appeared to be “more focused on addressing a perceived PR problem than the significant harms its products create.”
In brief
WiFi: Crew on a flight from Las Vegas turned off the wireless network after someone created a fake access point. The flight was full of people returning from the DEF CON hackers’ get together, so perhaps not surprising someone couldn’t resist the temptation to play around with their gadgets. It’s trivially easy to jam a network and set up a fake alternative, which is why we keep on saying it’s better to use mobile cellular data whenever you can. The Register
Photos: The FBI has advised against storing nude and sexually explicit content on cloud, email, or social media accounts, saying hackers are specifically looking for this type of content to extort victims. The agency lists typical tactics, including targeting passwords and PINs, impersonating customer service, and phishing messages. As always, never upload anything you wouldn’t be happy for the world to see.
Expansion: The volume of ransomware activity stayed roughly the same in the second quarter of this year, but the number of groups tracked by Check Point leapt from 71 to a record 93. One reason may be that it’s becoming much easier to set up a ransomware operation. A leak from one group’s infrastructure revealed a core team of just nine people, with most attacks carried out by “affiliates” using its platform. The group’s administrator built its management panel in about three days using AI coding assistants
Copyright: Journalists and civil society are increasingly being silenced by accusations of copyright infringement, according to Organised Crime and Corruption Reporting Project. The OCCRP says false reporting is part of an underground market in which “agencies” charge up to $5,000 to take down embarrassing content. This type of activity is “virtually invisible,” and leaves the burden of proof on victims to challenge accusations, the organisation warns.
Glasses: A University of Sydney study suggests that Meta’s smart glasses probably deserve their “pervert” tag. Building on reporting by the BBC, the research examines the ecosystem in which so-called “pickup artist” influencers use the glasses to record unwanted interactions with women in public and upload the results to social media. The paper argues that the issue is exacerbated by the glasses’ integrations with platforms such as Facebook and Instagram which enables content to be shared instantly. 404 Media
Musked: Since Elon Musk took over Twitter, the platform’s advertising revenue has fallen by two-thirds. Meanwhile, Lawfare says his AI-powered Wikipedia clone appears not to have updated any articles since April 24.
Updates
Patch Tuesday: Another flood of updates from manufacturers, as AI continues to aid vulnerability discovery. Microsoft released nearly 400 updates, fewer than last month’s 570 but double the number for June. 42 of the issues addressed by Microsoft are rated “critical,” and one affecting Windows is already being exploited.
Adobe: Updates to address more than 50 vulnerabilities, with critical issues flagged in ColdFusion, Campaign Classic.
Zoom: Fix for an issue that could have exposed anyone taking part in a meeting to attack. No interaction was required for an attacker to be able to run malicious code. We recommend checking Zoom is up-to-date before every meeting.
Signal: A new security feature is designed to verify that users are talking to the intended party. Automatic Key Verification ensures the correct association between a phone number, username, and its public encryption key. All verifications are done in the background, without any user interaction.